Вы находитесь на странице: 1из 13

Home Search Collections Journals About Contact us My IOPscience

Automatic recognition of defect areas on a semiconductor wafer using multiple scanning

electron microscope images

This content has been downloaded from IOPscience. Please scroll down to see the full text.

2009 Meas. Sci. Technol. 20 075503

(http://iopscience.iop.org/0957-0233/20/7/075503)

View the table of contents for this issue, or go to the journal homepage for more

Download details:

IP Address: 131.91.169.193
This content was downloaded on 30/08/2015 at 09:42

Please note that terms and conditions apply.


IOP PUBLISHING MEASUREMENT SCIENCE AND TECHNOLOGY
Meas. Sci. Technol. 20 (2009) 075503 (12pp) doi:10.1088/0957-0233/20/7/075503

Automatic recognition of defect areas


on a semiconductor wafer using multiple
scanning electron microscope images
Ryo Nakagaki1 , Toshifumi Honda1 and Koji Nakamae2
1
Production Engineering Research Laboratory, Hitachi, Ltd., 292 Yoshida-cho, Totsuka-ku,
Yokohama 244-0817, Japan
2
Department of Information Systems Engineering, Graduate School of Information Science and
Technology, Osaka University, 2-1 Yamada-Oka, Suita, Osaka 565-0871, Japan
E-mail: ryo.nakagaki.zs@hitachi.com, toshifumi.honda.td@hitachi.com and
nakamae@ist.osaka-u.ac.jp

Received 2 April 2009, in final form 6 May 2009


Published 10 June 2009
Online at stacks.iop.org/MST/20/075503

Abstract
A technique for high-precision and automatic recognition of defect areas on a semiconductor
wafer using scanning electron microscope (SEM) images is proposed. The proposed technique
inputs multiple SEM images formed by selectively detecting secondary electrons and
backscattered electrons emitted from the specimen by irradiating with primary electrons, and
defect areas are then automatically recognized by comparison with reference images. The
number of detected secondary electrons and backscattered electrons is highly dependent on the
surface roughness of the defect areas, namely the height and depth of defects; therefore, a
surface-roughness analysis from input images is conducted and the result is used to determine
the mixing proportion for multiple difference images. The proposed technique aims to obtain
high recognition accuracy for process wafers that contain various kinds of defects with a wide
variety of height and depth. The technique provides effective pre-processing for automating
the classification of defects, and is expected to contribute to improvements to the efficacy of
process monitoring and yield management in the fabrication of semiconductor devices.
Experimental results with two process wafers (involving 200 defect samples, each of which
belongs to one of the nine defect classes) have confirmed that the proposed technique is
capable of automatic recognition of defect areas with an accuracy of 98.9%.

Keywords: defect inspection, automatic recognition, defect classification, semiconductor


manufacturing, yield management
(Some figures in this article are in colour only in the electronic version)

1. Introduction detect, analyze and take appropriate countermeasures against


defects is of increasing importance for rapid yield ramping and
Developments to achieve ever-finer circuit patterns in assurance of stable manufacturing processes [2].
semiconductor devices are progressing on a continual basis Inline inspection systems, which are usually comprised
[1]. To reduce the pattern size, introduction of new processes of optical wafer inspection tools and scanning electron
and novel materials is essential, and, as a result, there has microscope (SEM)-based review tools, are deployed at
been an increasing diversity of defects that occur during the semiconductor wafer production sites for process monitoring
steps of device manufacturing. Furthermore, as the pattern [3]. Optical wafer inspection tools are used for the rapid
size is reduced, the size of defects that are critical to devices detection of defects on wafers. Such tools capture images of
has become smaller. Consequently, the ability to rapidly the wafer surface and compare them with those taken from

0957-0233/09/075503+12$30.00 1 © 2009 IOP Publishing Ltd Printed in the UK


Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

areas in the inspection image by comparing the two images.


Various kinds of defect characteristics are then quantitatively
calculated as a feature vector. The elements of the feature
vector include defect size, length, height, texture and so on.
Second, ADC executes classification processing for the given
feature vector. Several methods have been proposed for
classification processing. One is a rule-based classifier [4]
that classifies defects according to pre-determined rules. A
simple example of the rule-based classifier is thresholding
with respect to defect size, so that the defect classes are
defined by the size of the defects. In addition, a learning-based
classifier has been proposed [5–7], which involves classifying
the defects based on the statistical properties of feature vectors
and referencing learning data prepared from sample defects
collected in advance for each class. Recently, cases involving
the use of a classification algorithm called the support vector
machine (SVM) have also been reported [7]. However, the
Figure 1. Inline inspection system for semiconductor rule-based classification method, while offering the advantage
manufacturing. of not requiring the prior collection of samples, suffers from
a lack of flexibility with respect to changes in classification
criteria. On the other hand, the learning-based classification
scheme, while relatively amenable to changes in classification
criteria, suffers from a lack of stability in situations where
the collection of large amounts of sample data is difficult.
To circumvent these disadvantages, a hybrid-type approach,
which involves a combination of these methods, has also been
proposed [8].
Each of the above three steps of ADC needs to be robust
for various types of defects that occur in device manufacturing
process steps in order to obtain an accurate classification
Figure 2. ADC process flow.
result. This paper proposes a technique for the automatic
recognition of defect areas from an SEM image, which is the
neighboring chips that are formed in an array manner on first step of ADC. Defect area recognition processing serves as
a wafer. The pixel size of the images is in the order of a pre-processing for the calculation of feature vectors and the
several 100–10 nm, which is barely sufficient to determine classification of defects. Furthermore, the ability to identify
the presence of defects. The detected defects are subjected defect areas from given images with a high degree of precision
to review and analysis using a SEM-based review tool with is critical to obtain a classification result with high accuracy.
nanometer-order resolution. The defect review/analysis A large number of techniques have been proposed to
operation involves a manual review of the defect appearance, identify defects using comparison [9, 10]. However, although
elemental composition analysis and classification of the such techniques intend to rapidly identify the presence of
defects based on their size and type. The trend and frequency defects, they usually involve the setting of a coarse pixel size
of defect occurrence are then statistically analyzed from relative to defect size. For example, the technique in [9],
the review/analysis results (defect root cause analysis), and which is intended to detect defects from an SEM image, as in
the priority of countermeasures to be taken is determined the case of this study, is designed to detect defects measuring
(figure 1). approximately 30 nm from images with a 60 μm pixel size.
In order to improve the reliability and efficiency of the Reference [10], which reports a technique for wafer inspection
defect review/analysis process, it is necessary to test as many using dark-field optical images, is designed to detect 150 nm
wafers as possible and classify as many defects as possible defects from a dark-field image with a 2 μm pixel size. These
with a high degree of precision. In order to achieve this goal, defect recognition techniques are not suitable for ADC-type
there is an urgent requirement for SEM-based review tools defect analyses that require the acquisition of high-resolution
that automatically collect defect images in high throughput images of defects (for example, a 3 nm pixel size when dealing
and classify them with a high degree of precision [3]. with a 100 nm defect size).
Automatic defect classification (ADC) is a function that Recognition of hole regions from SEM images of contact-
automatically classifies defect images into pre-determined hole patterns has been reported as a defect recognition
defect classes based on their appearance [4–8]. Figure 2 and analysis technique using high-resolution images [12].
shows the typical flow of ADC processing. An image However, the target of inspection is limited to contact-hole
to be inspected and a reference image of a non-defective patterns, and the technique is not designed to address the
region are input for comparison. ADC first recognizes defect recognition of defect areas containing a wide variety of defects.

2
Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

(a) (b)

Figure 3. Characteristics of secondary and backscattered electrons. (a) Energy distribution of secondary and backscattered electrons. (b)
Emission of secondary and backscattered electrons from a circuit pattern with a trapezoidal-shaped cross section.

In other fields, such as thin film transistor (TFT) display with various energy levels. Figure 3(a), which represents
panel manufacturing, similar inspection techniques have also the relationship between the number of emitted electrons
been proposed [13, 14]. However, optical microscope images and the energy of electrons, indicates that large numbers of
are used for inspection, because the defects that occur in the electrons are generated, especially in two regions: the low-
fabrication of TFT devices are measured in micrometers, and energy region up to 50 eV and the high-energy region equal
techniques involving the use of SEM images have not been in intensity to the input energy (approximately 2 keV in the
proposed. figure). These electrons are referred to as secondary electrons
This paper proposes a defect area recognition technique and backscattered electrons, respectively.
applicable to the review/analysis of various types of defects The secondary and backscattered electrons exhibit
on semiconductor wafers through the use of an SEM with several different properties, for example their directionality.
nanometer-order resolution. Secondary electrons, because of their low energy, are easily
Section 2 discusses the principles and imaging amenable to a change in direction when the electromagnetic
characteristics of the defect review SEM. Section 3 describes field around the specimen is controlled, whereas it is difficult
the characteristics of semiconductor defects that are the objects to regulate the direction of backscattered electrons, because
of recognition and discusses the technological challenges of their high energy. For example, if an electron beam is
that must be overcome in order to achieve high precision irradiated on a circuit pattern with a trapezoid-shaped cross
and automatic recognition. Section 4 provides an overview section, as shown in figure 3(b), the secondary electrons can
of the proposed defect recognition technique that utilizes be detected above the sample by controlling the surrounding
multiple SEM images taken with different detectors, and electromagnetic field, regardless of whether they are generated
section 5 describes a defect surface-roughness analysis from the left or right inclines. In contrast, it is generally
method that is used to determine the mixing proportion for difficult to capture backscattered electrons emitted from the
multiple difference images in the proposed defect recognition left and right inclines using a single detector.
technique. Section 6 presents the experimental results that In addition, secondary and backscattered electrons differ
verify the effectiveness of the proposed technique. Finally, in the depth from which they are produced. For example,
section 7 provides a summary and possible expansion of future secondary electrons are generated from a region close to
research efforts. the surface (usually 10 nm) of the specimen. Therefore, if
the specimen surface contains a steep slope, the secondary
2. Defect review SEM electrons tend to exhibit the so-called edge effect, wherein
large numbers of electrons are generated near the slope. In
The principles of imaging with a defect review SEM and the contrast, for backscattered electrons, which are generated from
characteristics of images obtained are described with reference deeper sites, the edge effect is not observed. Furthermore,
to figures 3 and 4. Image capture is undertaken based on voltage contrast can be observed on secondary electron
the same principles as that used in a conventional SEM [15]. images, because the number of low-energy electrons detected
Specifically, an electron beam, emitted from an electron gun, by detectors has significant dependence on the surface
is focused using an objective lens and is directed onto the potential. In contrast, material contrast is highly visible
specimen. Electrons emitted from the surface of the specimen in backscattered electron images, because the number of
are detected by detectors. By converting the number of backscattered electrons is strongly dependent on the material
electrons detected (that is, the electron intensity) into a digital underlying the specimen.
value, a pixel value associated with the position of irradiation is By focusing on the differences in properties between
obtained. By controlling the deflector and causing the incident these two types of electrons, a suitable arrangement of
beam to scan the specimen surface in a rectangular pattern, a detectors and regulation of the electromagnetic field in the
two-dimensional digital image corresponding to the scanned specimen chamber enables the acquisition of a variety of
area is obtained. images exhibiting different properties. Figures 4(a) and (b)
Generally, irradiation of a specimen with an electron show a diagram of the review SEM and the detector layout,
beam causes the emission of a wide variety of electrons respectively. The review SEM is equipped with one detector

3
Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

(a) (b)

(c)

Figure 4. Review SEM used in the present study. (a) Diagram of the review SEM. (b) Layout of side detectors. (c) Example images
captured using the three detectors.

(referred to as the top detector) directly above the specimen


and two detectors (referred to as (left, right) side detectors)
diagonally over the specimen. The left and right detectors are
aligned at 45◦ to the chip layout coordinate system, as shown in
figure 4(b). This detector configuration is aimed at obtaining
high contrast images from circuit patterns, which usually run
horizontally or vertically to the chip layout.
Figure 4(c) shows examples of images captured using
the three detectors. While the image captured with the
top detector, which is subject to the edge effect, shows an
extremely bright circuit-pattern edge, the edge effect is absent
in the images captured with the side detectors; instead, a
Figure 5. Basic process flow of the comparison inspection
gradation of brightness corresponding to the orientation of technique.
the specimen surface is observed. The reason for this is that
the top and side detectors principally detect secondary and
backscattered electrons, respectively. Images acquired with
the top detector are referred to as top images and images
captured with the (left, right) side detectors are referred to as
(left, right) images.

3. Technical challenges for automatic defect area


recognition from SEM images

Semiconductor devices are formed in units of chips in an


array manner on a wafer. Therefore, inspection of a test
area is usually conducted by comparing it with the same
Figure 6. Top and left images of three types of defects.
area on an adjacent chip. The basic concept underlying this
approach, as detailed in figure 5, involves the following: image
registration to compensate for the difference in the field of Figure 6 shows the top and the (left) side images of three
view of two images, calculation of a difference image from types of defect (‘pattern open’, ‘foreign particle’ and ‘scratch’)
the two images and recognition of the defect area by execution that occurred on wafers during production. The pattern open
of thresholding on the resulting difference image. This basic defect is apparent in both the top and side images. On the other
concept is also employed in the proposed defect recognition hand, the foreign particle is visible in the top image, but is not
technique. However, the recognition of nanometer-order clearly visible in the side images. This is because the circuit
defects on semiconductor wafers requires solutions to the patterns near the defect prevent the emitted backscattered
following technical issues. electrons from reaching the detectors and that area becomes

4
Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

(a) (b)

Figure 7. Recognition results for two different particles using the conventional technique. (a) Top and left images of two types of particle
defects. (b) Recognized defect area using the conventional comparison technique depicted in figure 5.

a dead angle on the side images. For scratch defects, it is 4. Proposed defect recognition technique
evident that while they are clearly visible in the side images,
they are not visible in the top image, because only side images 4.1. Overall process flow
are suitable for detecting the surface roughness of a defect, Figure 8 provides an overall flowchart of the defect recognition
which is a distinctive feature of a scratch defect. algorithm employed in the proposed technique. The proposed
Therefore, images for recognition from suitable detectors technique consists of two steps: (A) defect recognition with
are highly dependent on the class of the defect. However, multiple SEM images and (B) defect surface analysis.
when recognizing defects, the defect class is unknown and In step (A), difference images of the inspection and
there is no way to designate suitably detected images. Thus, reference are calculated for each of the input images, and
for the automatic recognition of semiconductor defects, it is these results are then mixed and final defect detection results
necessary to explore techniques that utilize multiple images, are extracted by thresholding. The reason for conducting
comparison inspections of each input image is that the
on which the defect may or may not be visible.
algorithm must be robust for defects, which may or may not be
In addition, images that contain visual information of visible in any of the images. Step (A) uses a total of six images,
defects do not necessarily contain sufficient information for comprising both inspection and reference top images and (left,
defect recognition with high accuracy, which makes the task right) side images. Step (B) is performed to determine the
even more challenging. Figure 7 shows the top and the (left) mixing proportion of the three difference images calculated
side images of two different types of particle defects (surface in step (A). The surface analysis involves analysis of the
particle and embedded particle). These defects are visible roughness of the defect surface using four images, consisting
on both images; however, the images are not fully useful for of (left, right) side images of both the defect and the reference.
the recognition of defect areas. To illustrate the situation, The following section describes step (A), defect
detection experiments were conducted using the conventional recognition using multiple SEM images. The underlying
comparison inspection technique illustrated in figure 5. The principles and details of the algorithm used in step (B) are
results (figure 7(b)) indicate that, for an embedded particle, presented in section 5.
defect areas can be reasonably recognized using the side
images; however, the algorithm fails to detect the defect area 4.2. Defect recognition with multiple images
when using the top image. In contrast, for a surface particle, Figure 9 shows a flowchart for defect recognition with multiple
defect areas are reasonably recognized using the top image; images, which consists of the following steps.
however, when using the side images, the area surrounding
Step 1. Registration
the particle is also recognized as defect areas, which creates a
Registration is conducted in order to provide correction
problem of over-recognition. These results indicate that, for of the field of view deviation between the inspection and
embedded and surface particles, their respective top and side reference images. This process is further divided into the
images are not appropriate for automatic recognition with high following two steps.
accuracy.
Step 1-1. Global registration
Therefore, it is necessary to explore techniques that detect Global registration is first performed using the entire
defects among multiple images, on which the defect may or image. The purpose of the registration is to compensate for
may not be visible, and even if the defect is visible in one or the deviation of the field of view between the two images.
both of the images, the information may not be necessary and During the capture of inspection and reference images, wafer
sufficient for recognition with high accuracy. The following stage movement is required, which may cause a micrometer
section describes a defect recognition method capable of order positional error and this causes the deviation of the
addressing these challenges. field of view between the two images. A template-matching

5
Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

Figure 8. Overall process flow of the proposed defect recognition algorithm.

Figure 9. Process flow for defect recognition using multiple SEM images.

technique is applied, with the reference as the template and correlation coefficient is used for local registration. The
the inspection image as the target. To measure the degree number of small squares, which must be determined by the
of matching, a normalized cross correlation coefficient [17], degree of distortion in a given image, was set to 16 for
which is commonly employed for template-pattern matching the experimental tests described in section 6.
techniques, is used. Global and local registration processes for all three image
pairs (top, left and right image pairs) are not usually necessary,
Step 1-2. Local registration
as only one image pair is sufficient. For example, the top image
After global registration, local registration is performed.
pair is used for the calculation of the amount of positional shift,
Each of the inspection and reference images is divided
and the calculated values are used not only for the top image
into a large number of small square areas, and registration
pair, but also for the left and right image pairs, to create a
between corresponding small square regions of each image
deviation-corrected image.
is conducted. Local registration is required, because in an
electron microscope image, positional deviation of the beam Step 2. Difference image calculation
can arise from the electrostatic charge of the surface when A difference image is created by performing pixel-by-
the beam moves in a scanning motion, which results in image pixel difference computations on the registered inspection and
distortion. As with global registration, a normalized cross reference images. Circuit patterns formed on a semiconductor

6
Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

wafer are subject to fluctuations such as ‘line-edge roughness’,


and as a result these features in the inspection and reference
images are not perfectly identical. Moreover, positions at
which the electron beam is directed onto circuit patterns are
not perfectly identical from one pattern to another. Therefore,
if two images are compared in a simplistic pixel-by-pixel
manner, areas that should not be recognized as defects
can potentially produce a strong difference signal, resulting
in misidentification. To avoid this problem, perturbation
difference processing [16] is performed. The perturbation
method involves inclusion of the pixels surrounding a pixel
under inspection in the calculation of the difference value of
the pixel. By assigning the size of the adjacent pixels to be
considered as a parameter, an allowable pattern error between
the inspection and reference images is set. Figure 10. Recognition result for four different defects using the
proposed algorithm with weighting of ω = 1.0 and 0.0.
Step 3. Mixing the multiple difference images
Three difference images for the top image and the left and of the defects. The state of surface roughness is focused
right images obtained by step 2 are gathered into one difference on, because the surface-roughness state of defects, that is,
image by mixing them as shown in (1): their heights and depths, affect the manner in which secondary
Sub-total (x, y) = ω · |Sub-top (x, y)| and backscattered electrons are emitted, which determines the
+ (1 − ω) · (|Sub-left (x, y)| + |Sub-right (x, y)|)/2. (1) visibility of defects on both the top and the (left, right) side
images. A description of the underlying principles follows.
The values of Sub-total(x, y), Sub-top(x, y), Figure 10 shows the top and left images of four
Sub-left(x, y), Sub-right(x, y) represent the pixel val- representative defects in terms of different surface-roughness
ues of the mixed difference image, top difference image, left states (protrusion, dent, flat and bumpy (mixture of protrusion
difference image and right difference image at the coordinate and dent)). The results of defect recognition with mixing
position of (x, y), respectively. The weight coefficient ω is weight parameters of ω = 1.0 and ω = 0.0, using the
a parameter with the range of (0.0, 1.0). This represents a proposed algorithm (figure 8), are also shown. The recognition
relative weight between the top and side images with respect result is depicted as the boundary of a defect area, which is
to the difference image, and this coefficient is automatically calculated as a convex hull of the detected defect binary image.
set according to the method described in section 5. The subsections below discuss the relationship between the
Step 4. Thresholding recognition results and the surface-roughness state of defects.
Thresholding is performed on the mixed difference image. Discussion 1.
Pixels with intensities greater than a pre-set threshold value are For the case of protrusion-type defects, over-recognition
recognized as defects. of defect areas can arise when the side images are used. In
The purpose of the algorithm is to mix the three detected contrast, the defect can be properly recognized by using the top
images after calculation of the difference images, in order to image. The cause of over-recognition is assumed to be the fact
detect defects that are visible only on one of the three detected that the shadow (not shade) around the defect is recognized as
images without failure. The weight values for left and right a defect area.
difference images are treated as the same, because the left and
right images differ only in the direction of the detectors, and Discussion 2.
the imaging characteristics are assumed to be the same. Dent-type and bumpy-type defects are undetectable or
The following section explains the method for the under-recognized in the top images. It is necessary to use the
automatic determination of the weight parameters (ω) that side images for these types of defects, which require setting a
are required in the image mixing process. small value of ω. Even when the side images for recognition
are used, over-recognition does not occur.
5. Surface-roughness analysis for the determination Discussion 3.
of the mixing proportion for multiple difference For the case of flat-type defects, a real defect area is likely
images to be recognized, regardless of the value of ω.
Based on these discussions, consideration of the surface-
5.1. Key points
roughness states (protrusion, dent, flat and bumpy) seems to be
The defect recognition algorithm described in the preceding effective when setting the proportion value for mixing the three
section requires the setting of a proportion, in which the top, difference images. With regard to protrusion-type defects, the
left and right difference images are mixed. For this purpose, ratio for the top image is set large (e.g. ω = 0.8), which
a method of setting the mixing ratio was developed using aims at avoiding over-recognition caused by the shadow cast
the results from the analysis of the surface-roughness state by the defect. Conversely, dent- and bumpy-type defects are

7
Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

Figure 11. Process flow for the surface-roughness state analysis.

recognized by setting a small proportion ratio for the top image In this equation, Shade(x, y) and Tri(x, y) represent the
(e.g. ω = 0.2). In the case of flat-type defects, ω = 0.5 is set pixel values of the shading-enhanced and trinarized images on
as an in-between of the above two cases. the coordinate (x, y), respectively.
The following subsection describes the processing Examples of the results of trinarization are shown in
algorithm employed to categorize the defect into four types figure 11. White pixel areas indicate the uphill slant regions
of surface-roughness state. with pixel values in the shading-enhanced image that are
greater than THtri and the black pixel areas indicate the
5.2. Surface-roughness analysis algorithm downhill slant regions with pixel values that are smaller than
−THtri. The gray pixel areas indicate the flat regions.
Figure 11 shows the process flow for the surface-roughness
state analysis. First, two indices ((a) the defect height level Step 2. Extraction of defect shading
and (b) the degree of bumpiness) that express the state of Extraction of defect shading is performed by comparing
the surface roughness of a defect are calculated (steps 1–4), the trinarized shading-enhanced images with the inspection
and then the defect is categorized into four types of surface- and reference areas. Specifically, this is performed by
roughness state by thresholding (step 5). Detailed descriptions eliminating the uphill/downhill shading regions, which are
of each step are given below. present on both the inspection and reference areas, from
the trinarized shading-enhanced image of the inspection
Step 1. Shading enhancement and trinarization
area.
Difference computation of right and left inspection images
is performed to obtain the so-called shading-enhanced image. Step 3. Shading matching
This process is also performed for the reference image. In order to produce a shade map that evaluates the
Subtracting the right image from the left image produces an geometry of the uphill/downhill slant, a process called shading
image with enhanced shading and with a shading direction matching is performed. In this process, the relative position
that is the same as that in the left image. This means that of the white and black pixels in the trinarized shading-
for the case of a protrusion-type defect, an uphill slant (bright enhanced image is analyzed and output as a shading map
shading) emerges on the upper-left side of the defect and a (figure 12).
downhill slant (dark shading) emerges on the lower-right side First, the white pixel and black pixel images are extracted
of the defect, because the side detectors are aligned at 45◦ to from the trinarized image. Second, the shading map is
the chip layout, as shown in figure 4(b). calculated in order to examine the geometry of the uphill and
Next, trinarization of the shading-enhanced image is downhill slant. This is done by (a) shifting the black pixel
conducted to divide the image into three regions, namely uphill image to the white pixel image, (b) calculating the matching
slant, downhill slant and flat regions. In the shading-enhanced score at each shifted position and (c) creating the shading map
image, the uphill and downhill slant regions have positive and by arranging the matching score with respect to the shifted
negative values, respectively, and the flat regions have values position. The matching score at the position where the black
around 0. This leads to the use of the threshold THtri for pixel image is shifted by (X, Y ) to the white pixel image is
trinarization, as expressed in (2):
⎧ defined as (3)
⎨256 if Shade(x, y)  THtri
1
n
Tri (x, y) = 128 if − THtri < Shade(x, y) < THtri. Matching score = |ShdEnhncd(xi , yi )
⎩ n i=1
0 if Shade(x, y)  −THtri
(2) − ShdEnhncd(xi + X, yi + Y )|, (3)

8
Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

Figure 12. Algorithm for shading matching.


Figure 13. Examples of calculated shading maps.

where ShdEnhncd(xi , yi ) is the pixel value of the shading-


enhanced image at coordinate (xi , yi ), where the white and The degree of bumpiness is an index that quantizes the
black pixels are overlapped when the black pixel images are degree of coexistence of protrusion and dent regions. For
shifted by (X, Y ) to the white pixel image, and n is the number the case of uni-modal defects (protrusion or dent alone), the
of overlapped pixels. In other words, the matching score is shading map only has a value on the second or fourth quadrants,
the average of the absolute value of the difference of the pixel whereas the shading map of bumpy defects has values on
values of the shading-enhanced image when the white and both quadrants. The degree of bumpiness can be calculated
black pixels are overlapped. by evaluating the mixture ratio of the matching score in the
Figure 13 shows examples of the calculated shading maps. second and fourth quadrants of the shade map. Specifically,
The pixel value of the shading map image is normalized with the index is calculated according to (4), which uses the average
respect to the [0, 255] range for visibility. The center of the value of the shade map in the second (MapAveII) and fourth
shading map is the position where the shift value between black (MapAveIV) quadrants:
and white pixel images is (0, 0). For the case of a protrusion-
Degree of bumpiness =
type defect, the white pixels lie on the upper-left side of the ⎧  
black pixels, and this matches with the result that the region ⎪
⎪ |MapAveII − MapAveIV|
⎨ 100 × 1 −
with a large matching score lies in the upper-left area (area MapAveII + MapAveIV (4)
[II] depicted in figure 12). On the other hand, for the case of ⎪
⎪ if (MapAveII + MapAveIV = 0)

a dent-type defect, a matching score with a large value lies in 0 if (MapAveII + MapAveIV = 0).
the lower-right area (area [IV]). For the case of bumpy defects,
The value is normalized in the range of 0–100, as shown
a matching score with a large value likely lies on both areas
in (4). Defects with a value of 0 are considered to be uni-modal
[II] and [IV]. The difference of the surface state is expressed
and those with values around 100 are bumpy.
as the matching score and its position on the shading map, as
shown in figure 13. Step 5. Categorization
Thresholding is executed using the two calculated indices.
Step 4. Calculating two indices: the defect height level
Figure 14 shows the algorithm used for the categorization. A
and the degree of bumpiness
threshold Th bumpy is used to categorize bumpy-type defects
Two indices, the defect height level and the degree of
bumpiness, are calculated from the shading map. The height and two thresholds (Th protrusion, Th dent) are used to
level represents a quantization of the degree of the protrusion– categorize protrusion-type, flat-type and dent-type defects. In
flat–dent characteristics of the defect surface. The absolute the following section, experimental results for evaluation of
value and sign of the defect height level are defined as follows. the proposed algorithm are described.
The maximum shading matching score becomes the absolute
value of the height level and a positive sign is assigned when 6. Experimental results
the maximum score lies in the second quadrant (area [II]) and
a negative sign is set when the maximum score lies in the 6.1. Data
fourth quadrant (area [IV]). For the case where the maximum
score is 0, i.e. where either of the white and black pixels is not Defect SEM images obtained from two process wafers were
present in the trinarized shading image, the defect height level used for the experiments. The specifications of the data are
becomes 0. The maximum score lies in either area [II] or area shown in table 1. Data 1 contain images from a transistor
[IV], because the side detectors are aligned at 45◦ to the chip gate process, where the circuit pattern has a structure with
layout (figure 4(b)). When the absolute value of the defect height with respect to the base. Data 2 contain images from a
height level is greater than 100, the value 100 is assigned, Cu interconnect process, where the circuit pattern and the base
so that the range of the defect height level becomes [−100, have the same height level, because a chemical and mechanical
100]. polishing (CMP) process is applied to the surface of the wafer.

9
Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

Table 1. Details of the experimental data.

6.2. Surface-roughness analysis


Figure 15 shows scatter diagrams of the defect height level
and the degree of bumpiness index values calculated from
each dataset. Figure 15(a) shows the results from data 1.
Protrusion-type defects, such as surface particles and pattern
deformations, have a defect height level value over 10, and the
degree of bumpiness is less than 10 except for one example.
Scratches, which are of dent type or bumpy type, have a wide
distribution on both indices. The range of the defect height
level is from −90 to 0 and the range of the degree of bumpiness
is from 0 to 80. Scratches have a broad distribution on the
defect height level, because this class contains defects with a
wide variety of depth. The broad distribution on the degree of
bumpiness is because the scratch class contains defects with
both bumpy structures and simple uni-modal dent structures.
The proposed two indices successfully express the trend of
Figure 14. Algorithm for surface-roughness state categorization
using two indices. surface roughness of defects for data 1.
The same considerations are evident for data 2, with
distributions of the two indices shown in figure 15(b).
The total number of defect samples used for the experiments The surface particle protrusion-type defects have a defect
is 200 (data 1: 95, data 2: 105). height level over 10 and a degree of bumpiness less than
The defect classes included in data 1 are surface particle, 10. Conversely, scratches (dent-type defect) have a broad
pattern deformation and scratch. Surface particles and pattern distribution on both indices: ranging from −80 to −50 for
deformations are protrusion-type defects and scratches are the defect height level and from 0 to 100 for the degree of
dent-type or bumpy-type defects. The defect classes included bumpiness. Other dent-type defects, craters and pits, have
in data 2 are surface particle, scratch, crater, pit and pattern broad distributions, especially for the defect height level. The
deformation. Surface particles are a protrusion-type defect, scatter diagram also shows that pattern deformations (flat-type
and scratches, craters and pits are dent-type defects. Pattern defect) in the Cu interconnect have a value of almost 0 in terms
deformation is a flat-type defect. In total, nine defect classes of both indices.
are included in the experimental data. The proposed two indices have also successfully
Subsection 6.2 presents the evaluation results of the expressed the trends of surface roughness of defects for
surface-roughness analysis algorithm detailed in section 5 and data 2. This demonstrates that the four-type categorization
subsection 6.3 provides the evaluation results of the proposed algorithm (figure 14) is expected to provide appropriate results.
defect area recognition algorithm. The experiment described in the following subsection used

10
Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

(a) (b)

Figure 15. Scatter diagrams of the two calculated indices. (a) Data 1 and (b) data 2.

Figure 16. Examples of recognized defect areas (enveloped) of defect samples in data 1 and data 2.

the following values for thresholding: Th bumpy = 10, 6.3. Defect area recognition
Th protrusion = 5, Th dent = −1.
This subsection presents the results of the test of the proposed
The limitations of the proposed algorithm are given as
defect area recognition algorithm. Weight parameters, for
follows. In step 2 of the surface-roughness analysis algorithm
mixing of the multiple difference images, were set to ω =
(figure 11), shading information caused by non-defect regions
0.8 for protrusion-type defects, ω = 0.2 for dent- and bumpy-
is eliminated. This is simply accomplished by eliminating type defects, and ω = 0.5 for flat-type defects. The criteria
the shading region that occurs on both the inspection and for successful recognition are defined as the case where the
reference areas from the trinarized shading-enhanced image defect boundary calculated as the convex hull of the defect
of the inspection area. However, the algorithm does not work binary image contains a true defect region for approximately
appropriately for the case where a circuit pattern surrounding 90–110% of the area.
defect areas is deformed and the deformation causes Examples of the detected defect boundaries for data 1 and
the shading. The example expressed as 1 in figure 15(a) 2 are shown in figure 16. Each defect was properly recognized,
is the case of a surface particle. The calculated value of the regardless of the defect type, by setting the appropriate weight
degree of bumpiness was 22, which is much higher than the value for the surface-roughness state. The accuracy for
surface particles of the other samples. The trinarized shading recognition was 98.9% (94/95) and 99.9% (104/105) for data
image (figure 15(a)) shows that the deformation of a circuit 1 and 2, respectively, and 98.9% (198/200) for the total. Two
pattern (on the right side of the particle) causes the shading that examples that suffer from mischaracterization of the surface
was used in the shade matching process. This is an example roughness described in the previous subsection correspond to
of the coexistence of multi-type defects. The other example of the samples that were misrecognized. This shows that there is
a surface particle marked by 2 in figure 15(b) has a degree room for improvement in the accuracy of defect recognition
of bumpiness of approximately 10, which is larger than the by updating the surface-roughness analysis algorithm.
other samples of the same defect class. This result is caused
by the same reason as the first example. If the coexistence of 7. Summary
multi-type defects, as in data 1 and 2, is rare, then this may not
be an issue; however, this may be a problem when such cases A technique for automatic recognition of defect areas
are frequent. using SEM images was investigated with the objective of

11
Meas. Sci. Technol. 20 (2009) 075503 R Nakagaki et al

improving the efficiency of defect analysis in semiconductor [3] Guldi R L 2004 In-line defect reduction from a historical
manufacturing processes. The following is a summary of perspective and its implications for future integrated circuit
manufacturing IEEE Trans. Semi-Cond. Manuf.
conclusions.
17 629–40
• A novel defect recognition technique was proposed, in [4] Chou P, Rao A, Sturzenbecker M and Brecher V
which three types of SEM images are used; top, left and 1997 Automatic defect classification for semiconductor
manufacturing Machine Vis. Appl.
right SEM images are acquired by selectively detecting 9 201–14
the secondary and backscattered electrons emitted from a [5] Bennett M 1995 Automatic defect classification Proc. ASMC
given specimen. Conf. p 272
• The proposed technique is based on the comparison [6] Watanabe K, Takagi Y, Obara K, Okuda H, Nakagaki R
inspection technique. After registration of the inspection and Kurosaki T 2001 Efficient killer-defect control using
reliable high-throughput SEM-ADC Proc. ASMC Conf.
and reference images, difference images are calculated pp 219–22
for each detected image, and the results are mixed to [7] Chen S, Hu T, Liu G, Pu Z, Li M and Du L 2008 Defect
obtain a final difference image, followed by thresholding classification algorithm for IC photomask based on PCA
to obtain a final recognition result. The number and and SVM Congress on Image and Signal Processing vol 1
direction of secondary and backscattered electrons are pp 491–6
[8] Ritchison J, Ben-Porath A and Molasay E 2000 SEM based
highly dependent on the roughness of a defect surface; ADC evaluation and integration in an advanced process fab
therefore, the algorithm determines the proportion for Proc. SPIE 3998 258–68
mixing three difference images according to a defect [9] Hiroi T and Okuda H 2006 Robust defect detection system
surface-roughness analysis. using double reference image averaging for high throughput
• The surface-roughness analysis is achieved by analyzing SEM inspection tool Proc. ASMC Conf.
[10] Hamamatsu A, Shibuya H, Oshima Y, Maeda S, Nishiyama H
the shading information produced according to the and Noguchi M 2006 Statistical threshold method for
roughness state of each defect. The roughness semiconductor wafer inspection 12th Asia–Pacific Conf. on
characteristics are expressed by two indices: the defect NDT
height level and the degree of bumpiness. [11] Pepper D, Moreau O and Hennion G 2005 Inline automated
• Experimental tests were performed using 200 samples defect classification: a novel approach to defect
management Proc. ASMC Conf.
from nine defect classes of two process wafers, and a total [12] Takeda H, Sawai K, Uesugi K, Nakahara T, Mihara T,
recognition accuracy of 98.9% was verified. Nagaishi H and Sakurai K 2007 A new fast QC method for
testing contact hole roughness by defect review SEM image
For future work, evaluation tests will be conducted under a
analysis ISSM Conf.
volume production environment with a large number of defect [13] Boek S, Kim W, Koo T, Choi I and Park K 2004 Inspection of
samples. The adaptability of the algorithm to more advanced defect on LCD panel using polynominal approximation
device manufacturing processes that may contain a new type TENCON A 21–24 235–8
of defects will be verified. In addition, introduction of the [14] Jie C and Tsai D 2008 Independent component analysis-based
defect detection in patterned liquid crystal display surfaces
proposed approach to automatic defect classification will be
Image Vis. Comput. 26 955–70
used to verify the efficacy for automatic defect monitoring and [15] Goldstein J, Newbury D, Joy D, Lyman C, Echlin P, Lifshin E,
yield management. Sawyer L and Michael J 2003 Scanning Electron
Microscopy and X-Ray Microanalysis 3rd edn (New York:
References Kluwer Academic)
[16] Hiroi T, Maeda S, Kubota H, Watanabe K and Nakagawa Y
[1] ITRS 2009 International Technology Roadmap for 1994 Precise visual inspection for LSI wafer patterns using
Semiconductors 2008 UPDATE Overview p 5 subpixel alignment IEEE Workshop on Application of
(www.itrs.net) Computer Vision
[2] Braun A E 2007 Defect detection faces smaller, deadlier [17] Pratt K W 1991 Digital Image Processing 2nd edn (New York:
hurdles Semicond. Int. (4) Wiley)

12

Вам также может понравиться